An Ethical and Governance-Aware Data Collection Framework for Low-Resource Kazakh LLM Training

Abstract

The Kazakh language, spoken by over 13 million people, remains severely underrepresented in modern Large Language Model (LLM) research. A core reason is the absence of governance-aware, domain-specific training corpora. Many existing Kazakh NLP datasets rely on web-scraped or publicly available data, where consent, provenance, and demographic representation are not always explicitly documented. This paper proposes a six-layer interactive data collection framework for Kazakh LLM training that integrates ISACA COBIT 2019 governance controls and IEEE Ethically Aligned Design (EAD) v2 principles. The proposed architecture covers consent management, data processing, quality assurance, bias monitoring, and provenance tracking, together with a concrete set of proposed evaluation metrics. To the best of our knowledge, this is one of the first frameworks to integrate ISACA COBIT 2019 IEEE EAD principles into Kazakh language data collection for low-resource LLM training. We argue that governance and technical data quality are complementary objectives, and that the proposed framework addresses a documented gap in low-resource language AI infrastructure. Future work will implement and empirically validate the framework through a crowdsourced pilot study.

More by Zhanat Nurbekova

• Evaluating Large Language Models for Kazakh-Language Educational Assistants: Sentiment Classification and Question Answering with and Without ContextInformation · 2026
• A multifactor model for assessing programming knowledge based on the synchronization of AI analytics and expert evaluationFrontiers in Education · 2026
• Development of a data mining methodology for an analytical system supporting students’ professional developmentFrontiers in Education · 2026
• Development of a conceptual model of an intelligent system for forecasting career trajectories of students in technical specialtiesFrontiers in Education · 2026